PaperlayerView exact revision on arXiv

Source-linked AI summary

D-Scope: Decomposing and Steering Diffusion Transformers with Sparse Autoencoders

Xinyue Xu, Jiahao Zhang, Lijie Hu, Peter Hase, Hao Wang

arXiv:2609.39625v1cs.CVcs.AI

TL;DR

Fine-grained DiT edits need a way to select an SAE direction without per-feature descriptions or target-generation activations. D-Scope retrieves directions from visual evidence in activating patches and tests them through spatially masked interventions. Contrastive retrieval improves mean regional alignment over direct retrieval in evaluated configurations, but does not consistently improve outside-region preservation.

  • Problem

    Selecting an SAE decoder direction for a fine-grained edit has relied on per-feature textual descriptions or paired target-generation activations, adding information or processing beyond the source image.

    Source: §1 · p. 2

  • Method

    D-Scope pools SigLIP 2 embeddings of highly activating patches into visual centroids, matches target text to retrieve a feature, and applies its decoder direction within a spatial mask.

    Source: §1 · p. 2, §3.1 · p. 5

  • Results

    Contrastive retrieval yields higher mean regional target-alignment gains than direct retrieval across evaluated configurations, without consistently improving outside-region preservation.

    Source: §1 · p. 2, §5 · p. 10

  • Takeaways & Limitations

    High reconstruction fidelity alone does not characterize the visual evidence available for retrieval, and the selected direction’s semantic relevance does not guarantee outside-region preservation.

    Source: §4.4 · p. 9, §5 · p. 10

    Contrastive retrieval does not consistently improve outside-region preservation, and the readout results are a controlled demonstration rather than validation of a general-purpose content moderation system.

    Source: §4.4 · p. 9, ETHICS STATEMENT · p. 10

Abstract

from arXiv · show

Sparse autoencoders (SAEs) reveal visual structure in diffusion transformers (DiTs), but interpreting a feature does not establish whether it can be used to control generation. We introduce D-Scope (Diffusion Scope), a framework that connects feature interpretation to generation control through shared visual evidence. D-Scope aggregates SigLIP~2 embeddings of highly activating image patches into visual centroids. Matching target text descriptions against these visual centroids in the shared image-text embedding space then enables retrieval of individual features without per-feature text annotations. The underlying patches provide evidence for inspecting each selection, while spatially masked interventions test the corresponding decoder direction at varying strengths under fixed generation conditions. We characterize 150 SAEs across two model families and five layers, and introduce a benchmark of 100 target concepts with ten contexts each spanning under-specified and explicit-conflict conditions. Our empirical results show that high reconstruction fidelity can coexist with low dictionary utilization and limited visual-evidence coverage. Under per-case best-of-sweep strength selection, contrastive retrieval yields larger mean regional SigLIP~2 gains than direct retrieval across the tested steering configurations, without consistently improving outside-region preservation. D-Scope provides an inspectable framework for evaluating sparse DiT features through their visual evidence and the effects of their decoder directions on generation. The demo is available at https://jiahaozhang-public.github.io/d-scope/.

1 INTRODUCTION

D-Scope addresses how to select a decoder direction for a fine-grained edit from feature visual evidence, without per-feature text descriptions or target-generation activations. Its benchmark and dictionary analyses show both the promise and limits of that approach.

1 INTRODUCTION

  • Prior feature-selection approaches use per-feature textual descriptions or paired source-and-target activations, each requiring information or processing beyond the source image itself.Textual descriptions require annotation or generation; paired-activation selection requires a suitable target generation and its internal activations.Source: §1 · p. 2
  • The target is a fine-grained text-specified change that modifies a queried attribute while preserving other image content, using a pretrained SAE dictionary and its activating patches.Feature selection does not rely on per-feature textual descriptions or target-generation activations.Source: §1 · p. 2
  • D-Scope uses SigLIP 2 embeddings of highly activating patches to form visual centroids, matching target text directly or relative to a reference to retrieve a decoder direction.The selected direction is applied within a spatial mask while the source prompt, initial noise, and sampling settings remain fixed.Source: §1 · p. 2
  • High reconstruction fidelity can coexist with low dictionary utilization and limited visual-evidence coverage, so reconstruction alone does not establish how much retrieval evidence is available.Across evaluated configurations, contrastive retrieval has higher mean regional gains than direct retrieval, without consistently improving outside-region preservation.Source: §1 · p. 2
  • A single retrieved direction achieves up to 86% of Oracle Dense’s regional target-alignment gain under per-case best-of-sweep selection.The benchmark uses 1,000 fine-grained steering cases and evaluates regional target alignment and outside-region preservation under a fixed single-feature budget.Source: §1 · pp. 2–3

2 RELATED WORK

D-Scope relates sparse-feature interpretability to intervention evaluation and semantic retrieval. It indexes DiT SAE features through their activating image regions rather than relying on paired target-generation activations or direct text alignment of decoder directions.

2 RELATED WORK

  • D-Scope evaluates retrieved directions under a single-feature protocol because activation-based relevance alone does not establish steering effectiveness.It separately measures regional target alignment and outside-region preservation, following editing evaluations that distinguish edit fidelity from background preservation.Source: §2 · p. 3
  • Unlike methods that select features from paired source and target trajectories, D-Scope retrieves from text queries and activating patches without target-generation activations.RIEBench’s selection mechanism requires generating a target image and collecting its SAE activations before feature selection.Source: §2 · p. 3
  • D-Scope builds a visual index after DiT SAE training, avoiding per-feature text annotations, joint vision-language SAE training, and direct matching of decoder directions to text embeddings.Activation-weighted visual centroids support direct and contrastive matching; the selected feature identity links the centroid to its decoder direction.Source: §2 · p. 3

3 D-SCOPE

D-Scope links a feature’s visual evidence to its SAE decoder direction so text queries can retrieve a direction for a spatially masked intervention. The dictionary and visual index are built offline and reused, while retrieval and intervention are performed online.

3 D-SCOPE

  • D-Scope connects each feature’s activating patches and visual centroid to its decoder direction through the shared feature identity.The visual centroid supports semantic retrieval in the shared image-text embedding space; the corresponding decoder direction is used for intervention.Source: §3 · p. 3

3.3 SINGLE-FEATURE STEERING

  • The steering case holds the source prompt, initial noise, and sampling settings fixed while applying one retrieved direction within the target mask at every denoising step.The steering goal is to move the target region toward the query while preserving content outside the mask; strength is controlled by ρ.Source: §3 · p. 4, §3.3 · p. 6

3.1 SAE FEATURES AND VISUAL EVIDENCE

  • For each feature, the method selects at most one positive-activation patch per image and indexes it only if positive candidates occur in at least J distinct images.Eligible features retain their J highest-activation patches; the paper uses J = 64.Source: §3.1 · p. 5

3.2 VISUAL INDEX AND FEATURE RETRIEVAL

  • The visual index pools activating patches into a centroid, then retrieves one feature by matching a target query directly or relative to a generic-object or source-appearance reference.Direct matching can favor either the object or attribute in a query; contrastive matching is intended to emphasize the requested attribute or change.Source: §3.2 · p. 5

3.3 SINGLE-FEATURE STEERING

  • A retrieved feature’s decoder direction is adjusted for the SAE input setting before it is applied to the DiT residual stream.For LayerNorm inputs, the adjustment subtracts the decoder direction’s channel mean without inverting the input normalization.Source: §3.3 · pp. 5–6

4 EXPERIMENTS

The experiments assess which SAE features D-Scope can retrieve and how their decoder directions steer images. They show that dictionary reconstruction fidelity alone does not indicate retrieval readiness, and that retrieval strategy and dictionary choice shape target control and locality.

4.1 EXPERIMENTAL SETUP

  • The evaluation spans 150 SAEs, three generation environments, and a 1,000-case benchmark of 100 concepts across five attribute categories and ten source contexts per concept.Each case uses a fixed target-region mask; steering strength is selected to maximize Region SigLIP ∆, with LPIPS-out measured on the same output.Source: §4 · p. 6, §4.1 · pp. 6–7

4.2 SAE DICTIONARY QUALITY FOR VISUAL-EVIDENCE RETRIEVAL

  • An uncentered TopK SAE at Nitro layer 2 has EV 0.9997 but only 1.5% utilization and 0.5% visual-evidence coverage, showing reconstruction fidelity alone does not establish retrieval readiness.Centering generally increases utilization, while centered TopK has the highest median evidence coverage in the sweep.Source: §4.2 · p. 7, Figure 3 · p. 7

4.4 EFFECTS OF RETRIEVAL STRATEGY AND DICTIONARY CHOICE

  • Contrastive retrieval yields higher Region SigLIP ∆ than direct retrieval for all 15 dictionaries across three environments; mean gains rise from 21.8 to 30.6 on SANA-Sprint, 17.9 to 19.9 on Nitro-1-PixArt, and 19.8 to 28.4 on the SANA teacher (×10−3).The benchmark compares retrieval strategies across middle-layer dictionaries and reports Region SigLIP improvement and LPIPS-out.Source: Figure 4 · p. 8, §4.4 · p. 8, Table 1 · p. 9

4.3 SINGLE-FEATURE STEERING PERFORMANCE

  • Retrieved directions produce positive Region SigLIP ∆ values of 13.8–33.4 (×10−3; mean 23.1); contrastive retrieval averages 79%, 69%, and 82% of Oracle Dense gain across the three environments.The percentages are for SANA-Sprint, Nitro-1-PixArt, and the SANA teacher, respectively, averaged across 15 dictionaries.Source: §4.3 · p. 8, Figure 4 · p. 8

4.4 EFFECTS OF RETRIEVAL STRATEGY AND DICTIONARY CHOICE

  • Higher target alignment does not consistently improve outside-region preservation, which varies across generators.The paper distinguishes feature selection by activating visual evidence from locality, which depends on what the decoder direction writes back into the model.Source: §4.3 · p. 8, §4.4 · p. 9
  • Centering expands L1 and JumpReLU visual indices 1.8–4.4× and improves target alignment in 22 of 24 matched comparisons, while further TopK expansion does not consistently improve steering.Uncentered TopK already indexes more than 12,000 features.Source: §4.4 · p. 9

5 CONCLUSION

D-Scope retrieves SAE decoder directions from target text using visual evidence and evaluates their masked effects on generation. On the 1,000-case benchmark, one retrieved direction reaches up to 86% of Oracle Dense’s mean regional target-alignment gain; contrastive retrieval improves alignment, but not consistently outside-region preservation.

5 CONCLUSION

  • Up to 86% of Oracle Dense’s mean regional target-alignment gain is achieved by one retrieved direction; contrastive retrieval improves alignment but not consistently outside-region preservation.The comparison uses per-case best-of-sweep strength selection.Source: §5 · p. 10
  • High reconstruction fidelity can coexist with low dictionary utilization and limited visual-evidence coverage, so reconstruction alone does not characterize the evidence available for retrieval.Source: §5 · p. 10
  • Training choices shape the candidate feature pool, but a larger pool does not guarantee stronger steering.Source: §5 · p. 10
  • The nudity-related readout is a controlled demonstration, not validation of a general-purpose moderation system; broader use requires further evaluation and human oversight.The experiment uses images synthesized by the evaluated generator under controlled prompts.Source: ETHICS STATEMENT · p. 10

APPENDIX GUIDE

  • Table 2 summarizes practical DiT SAE recommendations, with steering findings scoped to middle-layer dictionaries and per-case best-of-sweep strengths.Source: APPENDIX GUIDE · p. 13, Table 2 · p. 14

A ADDITIONAL RELATED WORK

  • D-Scope indexes activating image patches to retrieve DiT features without per-feature text annotations or target-generation activations during selection.The related work contrasts this setup with feature selection using paired source and target trajectories.Source: A ADDITIONAL RELATED WORK · p. 15

B.1 TRAINING CONFIGURATIONS AND OPTIMIZATION

The training study compares 150 SAE dictionaries across two DiT generators, five layers, three SAE families, and five input settings. Dictionary evaluation measures reconstruction and feature use alongside the coherence and coverage of visual evidence.

B.1 TRAINING CONFIGURATIONS AND OPTIMIZATION

  • 150 dictionaries span two generators, five layers, three SAE families, and five input settings, with shared dictionary width, data splits, and optimization settings unless specified otherwise.Formal runs use a fresh initialization and one training seed.Source: §B.1 · p. 16, Table 3 · p. 17
  • Five input settings transform residual activations into SAE inputs; mean-centering and geometric-median settings initialize the shared decoder bias from the corresponding activation statistic.The input transform stays fixed during training, while the shared bias is optimized with the other SAE parameters.Source: §B.1 · p. 16
  • TopK retains the 64 largest rectified activations and adds an auxiliary reconstruction loss using eligible inactive features; L1 and JumpReLU use tuned sparsity coefficients.Source: §B.1 · p. 16

B.2 DICTIONARY EVALUATION

  • Dictionary evaluation measures explained variance for reconstruction fidelity, mean positive activations per token for realized sparsity, and training-window activity for utilization.The selected checkpoints are evaluated on the same 1,000 held-out prompts within each generator.Source: §B.2 · pp. 17–18
  • Visual coherence measures similarity among eligible features’ evidence patches, while coverage measures how broadly sufficient visual evidence is available across sampled features.Evidence patches are sampled from 64 distinct images per feature; coherence uses SigLIP 2 embeddings.Source: §B.2 · p. 18

B.3 ADDITIONAL ANALYSES OF DICTIONARY QUALITY

Dictionary quality varies across sparsity, coherence, evidence coverage, reconstruction, and utilization, so reconstruction alone is not a sufficient basis for selecting a dictionary or layer. Centered TopK performs strongly overall, but layer-specific results still matter.

B.3 ADDITIONAL ANALYSES OF DICTIONARY QUALITY

  • TopK achieves mean ℓ0 of 64 in every configuration, while L1 ranges from 56.4 to 121.4 active features and JumpReLU from 51.0 to 225.6.Achieved sparsity varies after full training despite a common pilot-selection target; family comparisons should account for realized sparsity.Source: §B.3 · p. 18, Figure 5 · p. 19
  • Mean-centered TopK and JumpReLU have median DINOv3 coherence of 0.225 and 0.234, but median coverage of 75.3% and 6.0%, respectively.Coherence measures consistency among eligible features, while coverage measures the fraction of sampled features with sufficient evidence.Source: §B.3 · p. 18, Figure 5 · p. 19
  • Centered TopK has median utilization of 98.6% and 98.9%, with median coverage of 75.3% and 72.0%, respectively, across the two centered recipes.At Nitro-1-PixArt layer 2, geometric-median-centered L1 reaches 12.5% coverage versus 0.5% for both centered TopK variants, supporting layer-specific evaluation.Source: §B.3 · p. 18
  • Across 30 matched pairs, layer 26 has lower SAE input-space EV and higher utilization in every pair; coverage rises in 29 and is unchanged in one.These endpoint results show that stronger reconstruction does not necessarily coincide with broader dictionary use or greater evidence availability.Source: §B.3 · pp. 18–19, Figure 6 · p. 20
  • For Nitro-1-PixArt mean-centered TopK, coverage falls from 75.5% at layer 8 to 52.2% at layer 14, then rises to 99.5% at layer 26.Intermediate-layer changes are not uniformly monotonic, motivating joint examination of reconstruction, utilization, and evidence coverage at each candidate layer.Source: §B.3 · p. 19

B.4 COMPLETE DICTIONARY EVALUATION RESULTS

  • Tables 4 and 5 report configurations for SANA-Sprint and Nitro-1-PixArt, including utilization, coverage, dictionary-level coherence, and unavailable or undefined evaluations.Recipe summaries and matched comparisons use unrounded values.Source: §B.4 · p. 19, Table 4 · p. 20, Table 5 · p. 22

C FEATURE CARDS AND VISUAL INDEX

The visual index is built from feature-activating catalog patches and supports retrieval without text annotations or access to steering outcomes. Index size tracks utilization, but does not consistently track target gains.

C FEATURE CARDS AND VISUAL INDEX

  • Each of 30 layer-14 dictionaries receives a catalog built without text annotations, benchmark prompts, or steering outcomes; each feature card uses its 64 highest-positive-activation images.One patch per image supplies the evidence, and those patches are pooled into the feature's visual centroid.Source: C FEATURE CARDS AND VISUAL INDEX · p. 23
  • A feature is indexed only when it has positive evidence in 64 distinct catalog images; 41.3% of all 552,960 features are indexed across dictionaries.Features with fewer positive images retain cards and centroids but are not retrieved.Source: C FEATURE CARDS AND VISUAL INDEX · pp. 23–24
  • The catalog index is less restrictive than coverage estimation: it uses all catalog images, whereas coverage samples 1,000 images and requires activation in at least 6.4%.The index-membership threshold is 0.64%; indexed share exceeds measured coverage for 25 of 30 dictionaries, with the other five within one standard error.Source: C FEATURE CARDS AND VISUAL INDEX · p. 24
  • Index size is almost perfectly rank-correlated with utilization (Spearman 0.99–1.00), but larger indices are not consistently associated with higher target gains.Among the remaining dictionaries, index-size correlations with target gain range from 0.14 to 0.69 on SANA and −0.76 to −0.48 on Nitro.Source: C FEATURE CARDS AND VISUAL INDEX · p. 24

D FINE-GRAINED STEERING BENCHMARK

  • Each benchmark target concept has five under-specified and five explicit-conflict contexts.The benchmark tables present the concept/context structure and matched rose-blooming cases.Source: Table 8 · p. 25, Table 9 · p. 25

D.1 BENCHMARK AND EVALUATION PROTOCOL

The evaluation tests retrieved SAE directions on a 1,000-case benchmark using fixed target regions and a calibrated strength sweep. It compares regional target alignment and outside-region preservation under per-case best-of-sweep selection.

Benchmark construction

  • The benchmark spans 100 concepts in five categories, each with five under-specified and five explicit-conflict contexts, for 1,000 cases.Cases specify a source prompt, target query, reference query, and target region; oracle references additionally use a full target prompt.Source: §D.1 · p. 25

Generators and dictionaries

  • Steering covers SANA-Sprint, Nitro-1-PixArt, and the SANA teacher at 1024×1024, intervening on the post-block residual stream at layer 14.The two student generators use 30 dictionaries total; the teacher reuses the SANA-Sprint dictionaries.Source: §D.1 · p. 25

Feature retrieval

  • Retrieval uses text queries and precomputed visual indices without target images or steering outcomes, and fixes the selected feature across strengths and seeds.Reference queries describe the generic target object for under-specified cases and the competing source attribute for explicit-conflict cases.Source: §D.1 · p. 25

Masked intervention

  • A source-image mask remains fixed within each case and seed, and masked residual updates use constant strength across denoising steps over a ten-value sweep.The sweep is calibrated separately by generation setting; LayerNorm subtracts the decoder atom's channel mean, while the other listed adjustments leave it unchanged.Source: §D.1 · p. 26

Evaluation metrics

  • Region SigLIP Δ measures the steered crop's target-query alignment change from the unmodified source; LPIPS-out measures outside-region change, with lower values indicating better preservation.The target crop expands the region bounding box by 15% per side; LPIPS-out excludes a 12-pixel dilation of the target region.Source: §D.1 · p. 26

Strength selection and aggregation

  • Strength is selected separately for each case and seed for each dictionary–retrieval combination, and reported metrics come from the same selected output.Three seeds are evaluated on common valid cases, with case averages summarized by the mean and sample standard deviation across seeds.Source: §D.1 · p. 27

D.2 ADDITIONAL ANALYSIS

Contrastive retrieval improves mean regional target alignment over direct retrieval across the evaluated dictionaries, but its effect on outside-region preservation varies. Category difficulty and student-to-teacher transfer further qualify the results.

Direct versus Contrastive retrieval

  • Across all three generation environments, Contrastive SAE retrieval increases mean Region SigLIP improvement over Direct SAE for all 15 evaluated dictionaries.LPIPS-out decreases for 8 of 15 SANA-Sprint and 14 of 15 Nitro dictionaries, but increases for all 15 SANA-teacher dictionaries.Source: §D.2 · p. 27

Concept category and source context

  • Color and texture or pattern are comparatively easy to steer, while local state or geometry is more difficult with a single SAE direction.Contrastive retrieval generally improves target alignment, with particularly pronounced explicit-conflict gains in the SANA environments.Source: §D.2 · p. 27

Student-to-teacher transfer

  • Mean-centered TopK with Contrastive SAE yields Region SigLIP improvements of 31.7 on the SANA student and 28.8 on the teacher, while LPIPS-out changes from 157.3 to 234.5.The teacher uses the same student-trained dictionaries, feature indices, and retrieved feature identities, but the control–preservation operating point changes.Source: §D.2 · p. 27

E ADDITIONAL APPLICATIONS OF LEARNED SAE FEATURES

Single SAE features serve as interfaces for both style suppression and semantic readout. Style suppression reduces target-style classifications under preservation constraints, while a nudity-associated feature provides a strong held-out readout.

E ADDITIONAL APPLICATIONS OF LEARNED SAE FEATURES

  • 53.8% for SANA and 44.0% for Nitro target-style suppression, versus 3.5% and 1.2% under mismatched interventions, with non-target styles and object content constrained.Suppression remains incomplete and varies substantially across styles.Source: E ADDITIONAL APPLICATIONS OF LEARNED… · p. 27

F.2 SINGLE-FEATURE SUPPRESSION PROTOCOL

  • The selected SAE code is rescaled during style suppression, unlike the unit-normalized decoder-direction edits used in the steering benchmark.For some preprocessing settings, the protocol restores fitted scale or recenters and normalizes the edited input; residual preservation applies before nonlinear restoration.Source: E ADDITIONAL APPLICATIONS OF LEARNED… · p. 27, §F.2 · p. 31

E ADDITIONAL APPLICATIONS OF LEARNED SAE FEATURES

  • A single frozen SAE feature selected on a labeled discovery set achieves AUROC 0.980 on 118 held-out images without fitting an additional probe.The held-out images use disjoint scenes and prompt-template families from discovery.Source: E ADDITIONAL APPLICATIONS OF LEARNED… · pp. 27–28

F.1 STYLE EVALUATOR ADAPTATION

  • Generator-specific style evaluators replace the released classifier, which achieves only 9.9% and 9.5% top-1 accuracy on the two generators’ external evaluation sets.The study retains ten styles with at least 80% per-class accuracy on validation and external images for both generators.Source: §F.1 · p. 29

F.2 SINGLE-FEATURE SUPPRESSION PROTOCOL

  • The suppression study evaluates 30 frozen layer-14 dictionaries across eight object contexts and five seeds, with baseline, targeted, and mismatched conditions sharing prompts, noise, and sampling settings.Target-style evaluation also uses generator-specific evaluators and measures retained-style and plain-object accuracy.Source: §F.2 · pp. 29, 31
  • Each target style receives one feature selected independently per dictionary using differences in normalized activation shares between that style and the remaining styles.The protocol fixes the selected feature, threshold, and scaling coefficient during generation, while re-encoding current residual activations at each denoising step.Source: §F.2 · pp. 29–30

F.3 QUANTITATIVE RESULTS

The style-suppression results show a trade-off between suppressing target styles and preserving other content, with substantial variation across styles. The accompanying readout protocol selects a feature using labeled discovery examples and evaluates it on held-out images.

F.3 QUANTITATIVE RESULTS

  • The selected configurations maximize mean UA subject to mean IRA and CRA of at least 90%, choosing unnormalized TopK for SANA-Sprint and mean-centered JumpReLU for Nitro-1-PixArt.Figure 9 summarizes target-style suppression, retained-style accuracy, and plain-prompt object accuracy across layer-14 dictionaries.Source: §F.3 · p. 31, Figure 9 · p. 32
  • Suppression varies substantially by style and generator: SANA strongly suppresses Mosaic but has little effect on Neon Lines, while Nitro shows the opposite pattern.The aggregate preservation constraint does not guarantee uniformly high retention for every target style.Source: §F.3 · p. 31, Table 14 · p. 33

F.4 QUALITATIVE EXAMPLES

  • Selected qualitative successes show visible attenuation, but examples also retain target-style attributes even when a classifier predicts a non-target style.The qualitative judgments are distinct from classifier-defined UA and do not establish complete visual removal.Source: §F.4 · p. 31, Figure 10 · p. 34, Figure 11 · p. 35

G.1 DATASET AND HUMAN ANNOTATION

  • The nudity readout dataset uses paired nude and non-nude variants across 40 scenes, split into disjoint discovery and test scenes and prompt-template families.Human labels follow realized image content; uncertain or invalid examples are excluded, leaving 77 discovery and 118 held-out test images.Source: §G.1 · p. 37, Table 15 · p. 37

G.2 LABEL-GUIDED FEATURE DISCOVERY

  • Feature discovery selects the SAE feature whose pooled generation-time activations best separate binary human labels on the discovery set.For a fixed K, the image-level score is the mean of the feature’s K largest spatial responses; the selected feature is evaluated on held-out images without an additional classifier or threshold.Source: §G.2 · p. 37

G.3 HELD-OUT READOUT AND FEATURE ANALYSIS

Feature 4867 achieves a high held-out nudity-readout AUROC, with little sensitivity to the tested spatial pooling sizes. Qualitative examples pair generated images with activation overlays, including matched nude-clothed scenes and hard negatives.

G.3 HELD-OUT READOUT AND FEATURE ANALYSIS

  • 0.980 AUROC on 118 held-out images for feature 4867, with AUROC varying only from 0.979 to 0.980 across pooling sizes K ∈ {4, 16, 64, 256, 1024}.The feature remains fixed for held-out evaluation; K = 16 is the reporting default.Source: §G.3 · pp. 37–38, Figure 13 · p. 38

G.4 QUALITATIVE READOUT EXAMPLES

  • Qualitative readout examples show feature-4867 activation overlays alongside generated images, including matched nude-clothed scene-and-seed pairs and potentially confounding non-nude examples.The hard-negative examples include content such as swimwear, sports clothing, face close-ups, and bare limbs.Source: §G.4 · p. 38, Figure 14 · p. 39, Figure 15 · p. 40

Related papers

Loading 2609.39625v1…